2026-08-12 09:59:19 +08:00
|
|
|
|
# Agent Note: 会话的 agent 由一份 preset cordis.yml 组装而成
|
2026-08-03 20:30:33 +08:00
|
|
|
|
|
|
|
|
|
|
Status: implemented
|
|
|
|
|
|
|
|
|
|
|
|
[English](2026-08-03-per-session-agent-presets.md) | 中文
|
|
|
|
|
|
|
|
|
|
|
|
## 问题
|
|
|
|
|
|
|
2026-08-05 16:40:03 +08:00
|
|
|
|
一个 `dsh` 进程服务多个会话，但决定 agent（智能体）究竟是什么的那套组装——它的工具、人设、提示词段落、委派后端——由启动器所引导的 `cordis.yml` 一次性固定给整个进程。若某个部署希望一个 benchmark 精简 agent 与一个完整编码 agent 并存，就必须跑两个进程；而现有的变通方案（`apps/cli/config/minimal.cordis.yml`，一个用来禁用工具行的 `--config` 覆盖层）会一次性改变所有会话。
|
2026-08-03 20:30:33 +08:00
|
|
|
|
|
2026-08-18 19:00:37 +08:00
|
|
|
|
对"让会话自选组装"最直觉的理解，是 loader 需要新增一层。其实不需要。[`dsh-tools`](../../../../packages/core/tools/README.zh.md) 与 [`dsh-system-prompt`](../../../../packages/core/system-prompt/README.zh.md) 本就按调用方上下文的 scope 分层归档注册，而且 [agent 本身就是一个注册 scope](2026-07-08-agent-scope-contexts.zh.md)。此前缺的只是一种把整份 `cordis.yml` 指向某一个 agent scope 的办法。
|
2026-08-03 20:30:33 +08:00
|
|
|
|
|
|
|
|
|
|
## 决策
|
|
|
|
|
|
|
|
|
|
|
|
**preset** 是一个目录，其中放置一份 `agent.cordis.yml`。agent 工厂的 `setup(agentCtx)` 把它作为 Cordis `include` 子树，挂载到该 agent 的 scope 上下文之下。entry 上下文沿原型链连到子树被挂载时所在的上下文，因此 preset 内部的每一次注册都落进该 agent 的分层，并随 agent 一起卸载。没有任何注册表新增分层，也没有任何已在运行的会话被触及。
|
|
|
|
|
|
|
|
|
|
|
|
组装划分为两个平面，依据是什么必须共享，而不是什么感觉上与 agent 有关：
|
|
|
|
|
|
|
|
|
|
|
|
| 平面 | 实例数 | 内容 |
|
|
|
|
|
|
|---|---|---|
|
2026-08-10 12:45:05 +08:00
|
|
|
|
| 宿主 | 一份 | 注册表本身（`tools`、`systemPrompt`、`agents`、`agent-loop`、`sessions`）、跨会话设施（持久化、查询、投影、存储、设置、凭据、遥测）、这些设施所解析的 subagent provider，以及 web 宿主 |
|
2026-08-04 12:46:57 +08:00
|
|
|
|
| agent | 每会话一份 | 单个 agent 对这些注册表的贡献：工具插件、人设与提示词段落、压缩策略 |
|
2026-08-03 20:30:33 +08:00
|
|
|
|
|
|
|
|
|
|
模型路由不进 preset。`installAgentLlmTarget` 已经是 provider、model 与 reasoning effort 的按 agent 可替换点；而挂在 preset 内部的 LLM 适配器永远不会被 `agent-loop` 解析到，因为后者位于宿主平面。
|
|
|
|
|
|
|
refactor(preset): bundle the shipped presets inside dsh-agent-presets
Review asked why the launcher special-cases one plugin's row. It no
longer does: the four shipped compositions move into the package
(presets/, in files), dsh-agent-presets resolves its own shipped root
and prepends it before configured roots (includeShippedRoot, default
true, opt-out for bare-machinery embedders), and the per-composition
derived patch, its spec, and the dump layer are deleted — profile-boot
and dump-config return to plain layer stacking. The always-load
guarantee now rides the schema default instead of patch ordering, so a
whole-config replacement keeps the shipped set and the squash, reload
freeze, and dump divergence stop being possible.
Gate globs, the web scaffold, and both preset browser lanes drop their
hand-fed shipped roots; the roster e2e keeps asserting configured roots
beside the shipped four against the built lib.
Fixes #2863.
2026-08-21 12:37:57 +08:00
|
|
|
|
部署交付哪些 preset，取决于 `packages/preset/agent-presets/presets/` 下有哪些目录；清单是那份目录列表，而不是在此另抄一份。
|
docs(agent-presets): bring the note and the architecture map up to what shipped
The Agent Note was written when only the seam existed and never caught up.
Rewritten in place, per the implemented-note contract, with the four facts the
later work established:
- a preset file is an INPUT: `EntryTree.write()` persists a tree whenever the
Loader thinks the config changed, and a self-disposing plugin is enough, so
the inherited behaviour truncates a shipped preset to `[]` the first time a
session ends
- a plugin that looks itself up in the global registry breaks inside a preset,
because `register()` files into the calling context's scope — the general
rule behind the `dsh-tool-skill` fix
- an entry-local `isolate` realm is invisible to the agent's own scope too, not
only to the host, which is what makes a preset's registry that agent's own
and also why a consumer left outside the group silently contributes nothing
- switching is blank-only, and why it swaps the subtree rather than the session
`docs/architecture.md` gains an Agent Presets section: the map has to carry a
new architectural concept or it is wrong, and the root layout gains the group.
Both budget ceilings are raised rather than the content cut. `AGENTS.md` sat at
1774/1775 — one word of room, already far under the 5% headroom the standard
asks for — so no group line could be added at all; `architecture.md` was in the
same shape. Raising restores headroom instead of encoding "the map may not grow".
2026-08-04 11:17:26 +08:00
|
|
|
|
|
2026-08-03 20:30:33 +08:00
|
|
|
|
挂载默认按会话进行。实测一份十二行组装每会话约 3ms、约 600KB，因此隔离比任何共享方案都更划算；而由用户或 agent 写出的 preset 也因此拥有尽可能小的影响面。确实自带昂贵单例的 preset，可以用 Cordis 自身的 `isolate` 词汇显式选择共享：命名 realm 的 label 是进程级全局的，因此两棵子树只要写同一个 label 就解析到同一个实例。
|
|
|
|
|
|
|
feat(agent-presets): make the default preset a user setting
`config.default` becomes the composition base of an `agent-presets` settings
namespace, so the user document layers over the deployment's engineering
default and a person can change which preset new sessions get without a
restart.
The value is read per resolution rather than snapshotted: a hot-reloaded
document takes effect on the next session created, and every running session
stays on the preset it was composed from — which is the same rule the
session-header guard enforces from the other side.
`resolve()` read `config.default` directly, which would have made the whole
setting inert; it now goes through `defaultId` like every other caller.
The write-protection test is rewritten against a temp profile root. It was
passing vacuously: the un-overridden Loader REWRITES the composition it read —
stamping `disabled: true` onto the self-disposing row — so the committed
fixture had been mutated by the very run that proved the bug, and every later
run compared against the damaged file and passed. Building the preset in a
temp directory makes the assertion immune to its own failure mode, and it now
fails with a visible `+ disabled: true` when the override is removed.
Review follow-ups on this layer. The exported schema is
`AgentPresetSettingsSchema`, symmetric with the `AgentPresetSettings`
interface it resolves and self-describing at an import site. The `session.create`
JSDoc promised "the deployment's default preset" for an omitted `agentPreset`,
which this layer makes false — it now names the effective default. The
constructor records why it does not use `installSettingsSection`: that helper
re-judges what a consumer DERIVED across attach and detach, and nothing here is
derived. The provider-unload test disposes the fiber `ctx.plugin()` handed back
instead of reaching into `ctx.reflect.store`, and the write-protection wait says
why slack is the right shape for an absence assertion.
The real composition covers the layering too. `apps/cli` boots the shipped
`cordis.yml`, stores `agent-presets.default`, and asserts an unnamed session
composes from it — the package suite proves the layering against a hand-built
context, this proves the roster and the settings provider are wired to each
other. That test also pins the settings row at a temp file: it defaulted to
`$DSH_HOME/settings.yaml`, so a developer's own stored default decided the
outcome of a file whose whole point is that only the shipped root does.
The Agent Note records the per-resolution read and its correspondence with the
session header, and the vacuous-test finding above.
2026-08-04 00:07:11 +08:00
|
|
|
|
未指名 preset 的会话拿到哪一个，是一项用户设置（`agent-presets.default`），叠在组装自身的 `default` 之上——后者成为 `base`。两层都需要：组装里的值是部署交付的东西，在完全没有 settings 提供方时也必须照常工作；而设置是让人不必去改一份可能并不属于自己的 `cordis.yml` 就能调整的东西。
|
|
|
|
|
|
|
2026-08-03 20:30:33 +08:00
|
|
|
|
## 后果
|
|
|
|
|
|
|
2026-08-22 13:10:23 +08:00
|
|
|
|
**有效默认值在每次解析时读取，绝不保存快照。** 缓存下来就需要一个 `watch` 订阅和一条重载路径才能保持诚实，而解析后的 scope 本来就会重读热重载过的文档。读穿也不只是省事，它让边界本身是对的：新值作用于**下一个新建的会话**，每个运行中的会话保持它被构建时的那份组装。这条不变量正是 session 日志从另一侧执行的同一条——header 记录会话**创建时**的 id，此后空白期的任何切换由 `agent-preset/selected` 事件记录，因此读取方解析的是两者之和（`resolveSessionPreset`）、绝不单看 header：恢复重建的是其历史所产出的那份组装而不是恢复时的部署默认值，冷读记录的 presenter 在那份组装的层里解析，网关也会拒绝把一个活着的会话收编到它当前运行的 preset 以外的 preset 之下。快照会让两者恰好在设置改变的那一刻各说各话。
|
feat(agent-presets): make the default preset a user setting
`config.default` becomes the composition base of an `agent-presets` settings
namespace, so the user document layers over the deployment's engineering
default and a person can change which preset new sessions get without a
restart.
The value is read per resolution rather than snapshotted: a hot-reloaded
document takes effect on the next session created, and every running session
stays on the preset it was composed from — which is the same rule the
session-header guard enforces from the other side.
`resolve()` read `config.default` directly, which would have made the whole
setting inert; it now goes through `defaultId` like every other caller.
The write-protection test is rewritten against a temp profile root. It was
passing vacuously: the un-overridden Loader REWRITES the composition it read —
stamping `disabled: true` onto the self-disposing row — so the committed
fixture had been mutated by the very run that proved the bug, and every later
run compared against the damaged file and passed. Building the preset in a
temp directory makes the assertion immune to its own failure mode, and it now
fails with a visible `+ disabled: true` when the override is removed.
Review follow-ups on this layer. The exported schema is
`AgentPresetSettingsSchema`, symmetric with the `AgentPresetSettings`
interface it resolves and self-describing at an import site. The `session.create`
JSDoc promised "the deployment's default preset" for an omitted `agentPreset`,
which this layer makes false — it now names the effective default. The
constructor records why it does not use `installSettingsSection`: that helper
re-judges what a consumer DERIVED across attach and detach, and nothing here is
derived. The provider-unload test disposes the fiber `ctx.plugin()` handed back
instead of reaching into `ctx.reflect.store`, and the write-protection wait says
why slack is the right shape for an absence assertion.
The real composition covers the layering too. `apps/cli` boots the shipped
`cordis.yml`, stores `agent-presets.default`, and asserts an unnamed session
composes from it — the package suite proves the layering against a hand-built
context, this proves the roster and the settings provider are wired to each
other. That test also pins the settings row at a temp file: it defaulted to
`$DSH_HOME/settings.yaml`, so a developer's own stored default decided the
outcome of a file whose whole point is that only the shipped root does.
The Agent Note records the per-resolution read and its correspondence with the
session header, and the vacuous-test finding above.
2026-08-04 00:07:11 +08:00
|
|
|
|
|
|
|
|
|
|
|
2026-08-03 20:30:33 +08:00
|
|
|
|
**直接挂载的子树对启动审计不可见。** 它不会把自己关联到 `Entry`，因此不在 `ctx.loader.entries()` 中，`assertEntriesActivated` 也看不到它。改由挂载过程自行校验各行，通过一个会公开自身 tree 的 `Include` 子类读取。
|
|
|
|
|
|
|
2026-08-06 11:57:00 +08:00
|
|
|
|
**preset 能写出 group，是因为 app 注册了它。** 跨行共享 realm 就是一个 `cordis:group` 行，而住在本工作区之外的 preset——也就是 Harness home 下由人或 agent 创作的那些，正是这套设计的目的——无法按名字解析 `@cordisjs/plugin-group`：Node 向上查找 `node_modules` 的路径从那里永远走不到 harness。因此 `boot()` 把 `cordis:group` 与 `cordis:include` 并排注册为 loader builtin，两者都经由环境模块管线加载，而不依赖被包含树自身的说明符解析。没有它，上文那套 `isolate` 词汇就只能一行一行地表达，提供方也永远无法与它的消费方归入同一组。
|
|
|
|
|
|
|
2026-08-03 20:30:33 +08:00
|
|
|
|
**preset 不得把服务发布进根 realm。** 这类服务是进程级全局而非按会话的，因此第二个挂载同一 preset 的会话会与第一个相撞——而这次相撞表现为 `setup` 永远观察不到的未处理 rejection，留下一个看起来健康、实则组装到一半的 agent。挂载改为直接拒绝它；本包的运行时不变量还会在每次服务通知时复查，因为从定时器或异步续体中发布的行会绕过一次性审计。
|
|
|
|
|
|
|
|
|
|
|
|
**失败会让 agent 回滚。** `setup` 在发布之前运行，因此挂载被拒绝会让 `ctx.agents.create()` 失败且不留残留。这正是 `setup` 是唯一受支持调用点的原因。
|
|
|
|
|
|
|
feat(agent-presets): make the default preset a user setting
`config.default` becomes the composition base of an `agent-presets` settings
namespace, so the user document layers over the deployment's engineering
default and a person can change which preset new sessions get without a
restart.
The value is read per resolution rather than snapshotted: a hot-reloaded
document takes effect on the next session created, and every running session
stays on the preset it was composed from — which is the same rule the
session-header guard enforces from the other side.
`resolve()` read `config.default` directly, which would have made the whole
setting inert; it now goes through `defaultId` like every other caller.
The write-protection test is rewritten against a temp profile root. It was
passing vacuously: the un-overridden Loader REWRITES the composition it read —
stamping `disabled: true` onto the self-disposing row — so the committed
fixture had been mutated by the very run that proved the bug, and every later
run compared against the damaged file and passed. Building the preset in a
temp directory makes the assertion immune to its own failure mode, and it now
fails with a visible `+ disabled: true` when the override is removed.
Review follow-ups on this layer. The exported schema is
`AgentPresetSettingsSchema`, symmetric with the `AgentPresetSettings`
interface it resolves and self-describing at an import site. The `session.create`
JSDoc promised "the deployment's default preset" for an omitted `agentPreset`,
which this layer makes false — it now names the effective default. The
constructor records why it does not use `installSettingsSection`: that helper
re-judges what a consumer DERIVED across attach and detach, and nothing here is
derived. The provider-unload test disposes the fiber `ctx.plugin()` handed back
instead of reaching into `ctx.reflect.store`, and the write-protection wait says
why slack is the right shape for an absence assertion.
The real composition covers the layering too. `apps/cli` boots the shipped
`cordis.yml`, stores `agent-presets.default`, and asserts an unnamed session
composes from it — the package suite proves the layering against a hand-built
context, this proves the roster and the settings provider are wired to each
other. That test also pins the settings row at a temp file: it defaulted to
`$DSH_HOME/settings.yaml`, so a developer's own stored default decided the
outcome of a file whose whole point is that only the shipped root does.
The Agent Note records the per-resolution read and its correspondence with the
session header, and the vacuous-test finding above.
2026-08-04 00:07:11 +08:00
|
|
|
|
**「preset 文件从不被回写」这条断言，必须先有失败的可能。** 最初那版在一次普通挂载之后断言文件未变，其实什么也抓不到：Loader 只在认定 config 变了时才会走到写路径，而那份组装里没有任何一行会自行销毁。回归用例改为植入一个自行销毁的行——真实 preset 在每次 agent 被拆除时都会命中的形状——并把组装放在临时根目录而不是 `fixtures/` 下：没有那个覆写，Loader 会回写它读入的文件，于是提交进仓库的 fixture 会被**恰恰是证明该缺陷的那次运行**改坏，之后每一次运行都拿改坏后的文件作比较从而通过。
|
|
|
|
|
|
|
2026-08-03 20:30:33 +08:00
|
|
|
|
**fiber 归属判定用对象同一性，而非 `uid`。** `uid` 是按 registry 计数的序号，因此两个不同根下的 fiber 会在它上面撞号；按 `uid` 比较曾导致一个运行时的子树为另一个运行时中发布的服务背锅。`ctx.plugin()` 返回的是 thenable 的 `Object.create(fiber)` 包装对象，与父链中出现的 fiber 永远不同一，因此子树在构造时捕获自己的 fiber。
|
|
|
|
|
|
|
docs(agent-presets): bring the note and the architecture map up to what shipped
The Agent Note was written when only the seam existed and never caught up.
Rewritten in place, per the implemented-note contract, with the four facts the
later work established:
- a preset file is an INPUT: `EntryTree.write()` persists a tree whenever the
Loader thinks the config changed, and a self-disposing plugin is enough, so
the inherited behaviour truncates a shipped preset to `[]` the first time a
session ends
- a plugin that looks itself up in the global registry breaks inside a preset,
because `register()` files into the calling context's scope — the general
rule behind the `dsh-tool-skill` fix
- an entry-local `isolate` realm is invisible to the agent's own scope too, not
only to the host, which is what makes a preset's registry that agent's own
and also why a consumer left outside the group silently contributes nothing
- switching is blank-only, and why it swaps the subtree rather than the session
`docs/architecture.md` gains an Agent Presets section: the map has to carry a
new architectural concept or it is wrong, and the root layout gains the group.
Both budget ceilings are raised rather than the content cut. `AGENTS.md` sat at
1774/1775 — one word of room, already far under the 5% headroom the standard
asks for — so no group line could be added at all; `architecture.md` was in the
same shape. Raising restores headroom instead of encoding "the map may not grow".
2026-08-04 11:17:26 +08:00
|
|
|
|
**preset 文件是输入，绝不是持久化目标。** 只要 loader 认为配置变了，`EntryTree.write()` 就会回写整棵树，而一个插件自我 dispose 就足以触发——销毁 agent 会 dispose 它的整棵子树。若继承该行为，它会重写自己读入的那份组装，实际后果是第一次会话结束时把随附 preset 截断成 `[]`。子树因此把 `write()` 覆盖为空操作。
|
|
|
|
|
|
|
|
|
|
|
|
**按自身名字回查全局注册表的插件，在 preset 里必然失效。** `ctx.tools.register()` 归档进**调用方**上下文的 scope，因此挂在 preset 里的插件只为一个 agent 注册，而不带 scope 的 `ctx.tools.get(name)` 理所当然查不到。`dsh-tool-skill` 正是这样写的，于是每次 preset 挂载都抛错；现在它与自己注册的那个定义比对。任何希望可被 preset 挂载的插件，都必须持有自己的注册对象，而不是按名字重新读取。
|
|
|
|
|
|
|
2026-08-12 09:59:19 +08:00
|
|
|
|
**entry 本地 `isolate` realm 不仅对宿主不可见，对 agent 自身的 scope 同样不可见。** 只有该组内部的行能解析到该服务。这正是让 preset 的 `skills` 注册表归属单个 agent 而非共享的原因——同时也意味着：留在提供方组之外的消费方会悄然解析到宿主注册表，然后什么都不贡献。
|
docs(agent-presets): bring the note and the architecture map up to what shipped
The Agent Note was written when only the seam existed and never caught up.
Rewritten in place, per the implemented-note contract, with the four facts the
later work established:
- a preset file is an INPUT: `EntryTree.write()` persists a tree whenever the
Loader thinks the config changed, and a self-disposing plugin is enough, so
the inherited behaviour truncates a shipped preset to `[]` the first time a
session ends
- a plugin that looks itself up in the global registry breaks inside a preset,
because `register()` files into the calling context's scope — the general
rule behind the `dsh-tool-skill` fix
- an entry-local `isolate` realm is invisible to the agent's own scope too, not
only to the host, which is what makes a preset's registry that agent's own
and also why a consumer left outside the group silently contributes nothing
- switching is blank-only, and why it swaps the subtree rather than the session
`docs/architecture.md` gains an Agent Presets section: the map has to carry a
new architectural concept or it is wrong, and the root layout gains the group.
Both budget ceilings are raised rather than the content cut. `AGENTS.md` sat at
1774/1775 — one word of room, already far under the 5% headroom the standard
asks for — so no group line could be added at all; `architecture.md` was in the
same shape. Raising restores headroom instead of encoding "the map may not grow".
2026-08-04 11:17:26 +08:00
|
|
|
|
|
|
|
|
|
|
**只有空白会话才允许切换。** 一旦跑过任何轮次，那段历史就是在该 preset 的工具下产生的，替换会留下无法执行的已记录 tool call，因此 `agentPreset.select` 返回 `agent-preset-locked`。空白期的切换保留 agent 与 session，只替换子树——因为宿主丢弃了它创建的 `AgentHandle`，也没有 delete RPC；而保留它们本身就是更好的结果，会话 id、workspace 挂接与 projections 都原地不动。该替换是"先卸后装"（两份组装会把同名工具注册进同一分层），因此它在拆除任何东西之前先解析新 preset，并在新组装装载失败时恢复原来的那一份。
|
|
|
|
|
|
|
2026-08-25 00:59:04 +08:00
|
|
|
|
**创作 preset 是一次 RPC，而且是特权 RPC。** 组装是一个文件，但“去文件系统里改它”并不是浏览器能提供的操作，因此名单在 `select` 之外新增了 `read`/`write`/`remove`。Connection 用一个浏览器会话认证创作操作、`list`、`select` 与完整 Host API：组装指明一个会话所运行的插件，因此读取它是侦察，写入它是任意能力；选择 preset 则没有授予 `session.create` 携带 `agentPreset` 时尚未拥有的能力。这份能力也不由 preset 授予：部署自带的默认 preset 本就带着 `bash` 与文件系统工具，因此任何被允许开启会话的调用方，早已能以本进程的身份执行命令。约束是 id 自身的性质（`[a-z0-9][a-z0-9-]*`），在它成为目录名之前就检查，而不是事后再去审视拼接出的路径；文本使用 loader 自身的 schema 与方言解析，因此保存不会留下任何会话都无法加载的文件。随部署提供的 preset 拒绝写入与删除，因为部署自带的那一份正是用来对照有问题的本地 preset 的——这也让“先复制、再编辑”成为创作路径本身，而非事后补充。
|
feat(web): author agent presets from a settings page
A composition is a file, but "edit it on the filesystem" is not a browser
affordance. The roster gains `read`/`write`/`remove` beside `select`, and
the browser gains a settings section over them: the presets as rows, one
composition open in a YAML editor at a time, and per-row default, duplicate,
and delete.
All four authoring methods are loopback-pinned. A composition names the
plugins a session runs, so reading one is reconnaissance, writing one is
arbitrary capability, and selecting one can move a session onto a preset
that edits the live runtime. `agentPreset.list` deliberately stays ordinary
and now reports `authorable`, so a surface knows whether creating is
possible at all rather than offering a button whose save always fails.
Authoring starts by duplicating: a shipped preset opens read-only because
the deployment's copy is what a broken local one is compared against. Ids
are contained before they become directory names, and the text is parsed
with the loader's own schema, so a save cannot leave a file no session
could load.
Fixes a defect the real-composition test found: a preset written under the
user's home could never mount, because the loader resolves a row against the
composition's own directory and Node's `node_modules` walk from there never
reaches the installed harness. The mount now records the host base and sends
bare specifiers there, leaving relative paths resolving from the preset.
Also closes the coverage the earlier surfaces in this stack shipped without —
the General row, the composer seat, and the plugin halves now have tests.
2026-08-04 12:23:40 +08:00
|
|
|
|
|
2026-08-18 19:00:37 +08:00
|
|
|
|
**在 agent 平面之外还有消费方的服务，不能搬进 preset。** 激进拆分把 `subagents` 注册表连同 spawn/fork 后端一起搬进了 delegation 组的 entry-local realm，于是 `dsh web` 直接起不来：`dsh-host-apiproxy` 是宿主行，它注入 `subagents` 来回答浏览器的跨会话查询（`listChildren`、`followup`），因而永远等待一个此刻只有会话才提供的服务。按会话各一份在两个层面上都是错的——provider 名只能注册一次，第二个会话本来也会相撞。注册表与所有共享后端，包括[固定的 Codex 与 Claude Code 产品 provider](2026-08-10-product-subagent-providers-in-shared-host.zh.md)，都属于宿主平面；preset 只贡献自己的 agent 应看见的委派**工具**，这些工具解析宿主注册表。`workflows` 保持 entry-local，因为 agent 之外没有任何东西读它。本该拦下它的是「检索注入方」这一步，而它没拦住：检索必须覆盖宿主包，而不只是 agent 平面的包。
|
2026-08-04 12:46:57 +08:00
|
|
|
|
|
|
|
|
|
|
**真实组装测试若禁用了某个宿主行，就无法审计该行。** web 组装测试把 `api-gateway`——也就是 api-proxy 本身——当作「有外部副作用的行」禁用了，而它恰恰是那个会以 pending 注入点名此次断裂的行。现在它在启用 api-proxy、并替换为 browse 目录选择器的前提下引导，启动审计因此覆盖整个宿主平面的注入图；只有端口、资源目录与遥测导出器仍然关闭。
|
|
|
|
|
|
|
feat(web): author agent presets from a settings page
A composition is a file, but "edit it on the filesystem" is not a browser
affordance. The roster gains `read`/`write`/`remove` beside `select`, and
the browser gains a settings section over them: the presets as rows, one
composition open in a YAML editor at a time, and per-row default, duplicate,
and delete.
All four authoring methods are loopback-pinned. A composition names the
plugins a session runs, so reading one is reconnaissance, writing one is
arbitrary capability, and selecting one can move a session onto a preset
that edits the live runtime. `agentPreset.list` deliberately stays ordinary
and now reports `authorable`, so a surface knows whether creating is
possible at all rather than offering a button whose save always fails.
Authoring starts by duplicating: a shipped preset opens read-only because
the deployment's copy is what a broken local one is compared against. Ids
are contained before they become directory names, and the text is parsed
with the loader's own schema, so a save cannot leave a file no session
could load.
Fixes a defect the real-composition test found: a preset written under the
user's home could never mount, because the loader resolves a row against the
composition's own directory and Node's `node_modules` walk from there never
reaches the installed harness. The mount now records the host base and sends
bare specifiers there, leaving relative paths resolving from the preset.
Also closes the coverage the earlier surfaces in this stack shipped without —
the General row, the composer seat, and the plugin halves now have tests.
2026-08-04 12:23:40 +08:00
|
|
|
|
**preset 的包名必须从 harness 解析，而非从 preset 解析。** `EntryTree.import()` 按行所属树的 `baseUrl` 解析，而 `Include` 把它设为组装文件所在的目录。这对相对标识符是对的，对包名却是致命的：本地创作的 preset 位于用户主目录之下，Node 向上查找 `node_modules` 永远够不到已安装的 harness，因此每一个 `@deepseek-ai/dsh-*` 行都会导入失败，整个 preset 无法挂载。随部署提供的 preset 掩盖了这一点——它们本就在安装目录之内。挂载在插入子树之前先记录宿主组装的基址，并把裸标识符送往那里，同时让相对路径继续从 preset 解析，使它自带的文件仍随它一同迁移。发现它的正是那个把 preset 写入临时根目录的真实组装测试。
|
|
|
|
|
|
|
docs(agent-presets): bring the note and the architecture map up to what shipped
The Agent Note was written when only the seam existed and never caught up.
Rewritten in place, per the implemented-note contract, with the four facts the
later work established:
- a preset file is an INPUT: `EntryTree.write()` persists a tree whenever the
Loader thinks the config changed, and a self-disposing plugin is enough, so
the inherited behaviour truncates a shipped preset to `[]` the first time a
session ends
- a plugin that looks itself up in the global registry breaks inside a preset,
because `register()` files into the calling context's scope — the general
rule behind the `dsh-tool-skill` fix
- an entry-local `isolate` realm is invisible to the agent's own scope too, not
only to the host, which is what makes a preset's registry that agent's own
and also why a consumer left outside the group silently contributes nothing
- switching is blank-only, and why it swaps the subtree rather than the session
`docs/architecture.md` gains an Agent Presets section: the map has to carry a
new architectural concept or it is wrong, and the root layout gains the group.
Both budget ceilings are raised rather than the content cut. `AGENTS.md` sat at
1774/1775 — one word of room, already far under the 5% headroom the standard
asks for — so no group line could be added at all; `architecture.md` was in the
same shape. Raising restores headroom instead of encoding "the map may not grow".
2026-08-04 11:17:26 +08:00
|
|
|
|
**preset id 对模型可见，必须写入日志。** 它决定工具集与提示词，因此被恢复的会话必须还原同一份组装；记录它属于会话事实，而非运行时状态。它与 `cwd` 并列写在会话头部，并由会话摘要携带，使选择器显示的是某个会话实际运行的 preset，而非部署当前的默认值。
|
2026-08-03 20:30:33 +08:00
|
|
|
|
|
feat(web): choose the agent preset on the new-session screen
The composer seat spent nearly all its life disabled: a session's
composition is fixed once a turn has run. Move the choice to the
new-session screen beside the workspace picker, where it still works,
and let the session header report what a running session runs.
The hero pick is staged rather than applied — that screen precedes the
session it belongs to. It lands when a session becomes current and is
still blank, which covers both the session a workspace connect creates
and the blank one it reuses; riding `sessions.create` would miss the
second. It is spent on first use, matching the workspace picker.
Fix the durability the header field claimed but never had: `agentPreset`
was declared on `SessionHeader` and dropped by the JSONL header line, the
SQLite sessions row, the derived query index, and the cold list
projection, so every resumed session came back composed from nothing.
Add the web e2e lane that would have caught it — the one lane that mounts
the shipped roster, which needed `cordis:group` in the scaffold's Loader
builtins, as `mountRootInclude` already registers.
2026-08-05 18:51:11 +08:00
|
|
|
|
**持久化的头部字段，在每个后端都写入之前都算不上持久。** `agentPreset` 带着正确的理由落在了 `SessionHeader` 上，而两个持久化后端都没有携带它：JSONL 头部行、SQLite `sessions` 行、以及派生的查询索引各自逐列映射头部，于是被恢复的会话回来时没有 preset，所有据以命名它的表层随之失声。`summarizeCold` 是同一个形状——它手工拼装冷列表行，而没有复用共享的投影。声明为持久的字段，需要一个跨越真实存储的测试，而不只是声明它的那个类型。
|
|
|
|
|
|
|
|
|
|
|
|
**这个选择属于它仍然可用的那个界面。** composer 座位几乎一生都处于禁用状态，因为一旦跑过一个轮次，preset 即固定。它移到了新建会话界面、工作区选择器旁边，选择在那里是**暂存**的：该界面先于它要应用到的会话存在，暂存值在某个会话成为当前会话且仍为空白时落地——这既覆盖工作区连接新建的会话，也覆盖它复用的那个空白会话，而搭 `sessions.create` 的便车会漏掉后者。它一经使用即被清空，与旁边的工作区选择器一致。至于运行中的会话在跑什么，则是其标题旁的一个只读标签：在那里放控件，等于承诺一次宿主会断然拒绝的切换。
|
|
|
|
|
|
|
2026-08-06 11:38:51 +08:00
|
|
|
|
**preset 放大的是宿主本来就在付的代价：没有任何东西会 dispose 一个 agent。** 用 `--expose-gc` 对随附组装实测：一个存活的 agent 在 `minimal` 上约占 0.17 MB、在 `standard`/`cordis` 上约 1.31 MB，挂载耗时分别约 38 ms 与 135 ms；进程里第一个 agent 另需约 7 MB，那是 Node 首次 import 模块的一次性成本，此后每次挂载共享。增长严格线性——10、30、50 个的单个增量一致——且 dispose 后基本全额回收（50 个 `standard` 占住 57.8 MB，释放后全部归还）。所以对象图并不泄漏，缺的是生命周期。`dsh-host-apiproxy` 创建后直接丢弃 `AgentHandle`，`archiveSession` 只改工作区注册表，`AgentRegistry` 没有驱逐机制，而宿主里唯一一处 dispose 是 JSON-RPC 服务器自身的关停。于是一个 web 宿主会留住它接触过的每一个会话，组装 preset 之后每个约 1.3 MB，而在此之前约 0.2 MB。注意：剪枝挂载注册表在这里没有用——它丢弃的是 fiber `uid` 已清空的记录，而永不死亡的 agent 永远不会清空它。
|
|
|
|
|
|
|
|
|
|
|
|
- 遗留 TODO：idle agent 驱逐——会话持久化后 dispose，恢复时重新挂载。它属于持有 handle 的那个宿主，不属于本 seam。
|
|
|
|
|
|
|
2026-08-03 20:30:33 +08:00
|
|
|
|
## 考虑过的替代方案
|
|
|
|
|
|
|
|
|
|
|
|
**在 scope 注册表中新增 preset 分层。** `ScopedLayers.merge()` 把全局层与恰好一个精确 scope 层合并。新增中间层可以让多个会话共用一份已挂载的组装，但它要改动 `dsh-scope` 及每个 scope 感知的注册表，换来的只是毫秒级的开销节省，而且会让 preset 的注册获得一个没有任何 agent 拥有的生命周期。
|
|
|
|
|
|
|
|
|
|
|
|
**把 agent 的 scope 键设为 preset。** 同一 preset 上的会话就能免费共享一层，但按 agent 的注册——`installAgentLlmTarget`、按 agent 的工具限制——会跨会话相撞。
|
|
|
|
|
|
|
2026-08-18 19:00:37 +08:00
|
|
|
|
**把每个 preset 作为子进程运行。** [`subagent-dsh-sdk`](../../../../packages/subagent/subagent-dsh-sdk/README.zh.md) 已经证明完整的子 harness 可行，隔离性也会是绝对的。但这同时意味着要按会话代理流式输出、审批与投影，那是一个传输层项目，而非组装问题。
|
2026-08-10 12:45:05 +08:00
|
|
|
|
|
|
|
|
|
|
**给产品 subagent 增加全局启用设置与独立设置页。** 进程级值会与 preset 争夺模型可见工具的所有权，也无法表达两个会话使用不同组装。产品 provider 留在宿主，普通 preset 行分别暴露 Codex 与 Claude Code 工具。
|
|
|
|
|
|
|
|
|
|
|
|
**为 Codex 与 Claude Code 的每种组合交付一份 preset。** 四个身份会复制完整 preset 组装，只为表示两条独立行。复制后的 preset 已能直接启用任一行，因此组合 preset 只增加名单与维护成本，不增加用户结果。
|